Clinical severity predominates over routine inflammatory and metabolic biomarkers in predicting hemorrhagic transformation after intravenous thrombolysis: internal temporal validation of incremental biomarker value

Background and Objective Hemorrhagic transformation (HT) remains an important safety concern after intravenous thrombolysis for acute ischemic stroke. We evaluated whether routine inflammatory and metabolic biomarkers improve HT prediction beyond clinical features. Methods We analyzed 847 thrombolysis-treated admissions from 2020 to 2026. In-hospital imaging-confirmed HT was the primary outcome, and strict 7-day ECASS II-compatible symptomatic intracranial hemorrhage (sICH) was exploratory. Admission date defined derivation ( n = 593; 67 HT events) and internal temporal-validation ( n = 254; 27 HT events) cohorts. The six-variable clinical model included age, baseline NIHSS, atrial fibrillation, large-vessel occlusion, the rapid-glucose field, and onset-to-treatment time. Models added ln(NLR), AIP plus residual cholesterol, or all three biomarkers. Twenty imputed datasets were generated using derivation-fitted imputation models. We assessed discrimination, calibration, Brier score, decision curves, and sensitivity analyses. Results HT occurred during 94 of 847 admissions (11.1%). The clinical model yielded AUCs of 0.812 (95% CI 0.761–0.860) in derivation and 0.801 (0.707–0.881) in temporal validation. Adding ln(NLR) increased the validation AUC to 0.818 (ΔAUC +0.017; DeLong p = 0.220). Adding AIP and residual cholesterol yielded an AUC of 0.803 (ΔAUC +0.002; p = 0.818). The full model yielded AUCs of 0.817 in derivation and 0.818 in validation (ΔAUC +0.018; p = 0.346). Baseline NIHSS was independently associated with HT (OR 1.084 per point, 95% CI 1.044–1.126; p < 0.001). Atrial fibrillation was independently associated with HT (OR 2.66, 95% CI 1.34–5.27; p = 0.005). The biomarkers ln(NLR), AIP, and residual cholesterol were not independently associated with HT. Validation calibration was similar, with a Brier score of 0.081 for both models. Biomarker gains were inconsistent across sensitivity analyses. In the exploratory sICH analysis (37 events), ln(NLR) was associated with sICH (OR 1.89, 1.04–3.42; p = 0.037). Validation included only 12 events, and the AUC increment was not statistically significant ( p = 0.097). Conclusion Clinical severity and established clinical features provided most of the predictive information for post-thrombolysis HT. Routine ln(NLR), AIP, and residual cholesterol did not provide sufficiently robust incremental predictive value beyond the clinical model. The NLR–sICH signal remains exploratory and requires external validation.